Question Intel or AMD for specific type of desktop

[cgeorgy]

Hello,
I'm having a hard time deciding on a desktop configuration.
Last time when I did this was more than 10 years ago, I have the feeling this is a more complex task in present.

My kind of workload : run 2-3 VMs in parallel, software development, Docker, various databases.
Display : I already have two monitors 2560x1440 at 60Hz with DisplayPort 1.2 (in/out) and USB Type-C port, I want them run in parallel with the new desktop.
Mobo size : should be Micro-ATX so that I can have more than 64 GB RAM and a case that is not too large ; 2x M.2 NVMe SSDs.
Things with moving parts : low noise is very important. I don't want cheap case, PSU, cooler and fans.
Games : only Starcraft 2, I am very happy how it runs on Asus ROG laptop from 2015 with i7 4710HQ, 16GB RAM and Nvidia GeForce GTX 860M ... just to have an idea about my gaming needs.

For max budget of USD 500-550 these three CPUs are most interesting for me : i9-10900K (10 cores) vs. Ryzen 5 5600x (6 cores) vs. Ryzen 7 5800x (8 cores)

1. For the workload described above, I have the feeling that more cores is better, even if the CPU with more cores is not the best in benchmarks.
Can I assume this ?
( Most people around urge me to buy Ryzen... without asking me why do i need a new computer )

2. I've read some articles stating that some i9 models have slightly better RAM latency compared with similar (in terms of pricing) Ryzen.
Is this right ?

3. If I were to choose i9 with embedded UHD Graphics 630, on a decent mobo and with no discrete graphics card, I am going to have a noticeable performance penalty ?
Should be possible to have 2x 2560x1440 at 60Hz on a single DisplayPort delivered by mobo ?

Thank you for help !

 

[damian101]

I'd go with some third generation Ryzen CPU, more cores for the money. Also very efficient under load (easier to keep the noise low).

 

[damian101]

Oh, and yes, Ryzen CPUs do generally have higher RAM latencies than Intel alternatives. That's why manually tightening the RAM latencies has bigger effect on Ryzen CPUs than on modern Intel CPUs. Another RAM-related difference between Intel and AMD is that on Ryzen higher RAM clocks reduces cache latencies as RAM clock is coupled with the Infinity fabric clock by default.
More memory ranks also increases load on the memory controller which can results in lower achievable RAM clocks and higher RAM latencies. But in practice having 2 ranks per channel decreases RAM latencies under load due to parallelization of memory requests, and at least with Ryzen 4 ranks per channel typically decreases RAM latencies under load even further, resulting in higher CPU performance.
I'd go with two dual-rank 32GiB modules for a 64GiB system. I think all consumer 32GiB modules are dual rank anyway though. 4 dual-rank 16GiB modules might give better performance at the same clock speed but at least on Ryzen it probably won't clock as high, only 2667 MHz is guaranteed for such a configuration.

 

[Shmee]

If you can wait a bit to build, I would snag a 5900X when available at MSRP. Great performance in both ST and MT apps.

 

[DrMrLordX]

@cgeorgy

You're looking to build what is basically a SFF PC and you want it low noise, but you're going to be potentially throwing some pretty heavy workloads at these proposed CPUs which may make them get a little toasty. The 10900k in particular could potentially eat up to 250W all by itself (albeit not for sustained operation; after 56s it's supposed to drop to 125W though some motherboards don't enforce that). The 5800X is going to top out at 142W sustained and the 5600X 88W. Of the three CPUs listed, the 5600X will easily be the easiest CPU to cool in your chosen enclosure. Are you sure 6c/12t will be enough for all your VMs? How many cores do you need per VM?

If you're going to step up to a 142W PPT CPU like a 5800x and if you want lots of cores for VMs, you really need to take the time to up your budget and get a 5900X or 5950X. The Intel chip will get you more cores than the 5600X and its availability is higher, but with its power-consumption you'll have to do a lot of work to either cool it or undervolt it.

 

[moinmoin]

1. For the workload described above, I have the feeling that more cores is better, even if the CPU with more cores is not the best in benchmarks.
Can I assume this ?
( Most people around urge me to buy Ryzen... without asking me why do i need a new computer )

When you mention VMs, dockers, multiple databases everybody is going to assume you are better off with more cores. On standard desktop iCore chips peak at 8-10 cores depending on gen, while Ryzen goes up to 16 cores. You could also go with HEDT and get a Threadripper with up to 64 cores.

Honestly if I were you I'd first try to better define your workload: How many threads is your workload typically running at any given time? How much of it is limited by ST performance? Based on that it's much easier to say what would be the minimum you should get.

 

[cgeorgy]

Oh, and yes, Ryzen CPUs do generally have higher RAM latencies than Intel alternatives. That's why manually tightening the RAM latencies has bigger effect on Ryzen CPUs than on modern Intel CPUs. Another RAM-related difference between Intel and AMD is that on Ryzen higher RAM clocks reduces cache latencies as RAM clock is coupled with the Infinity fabric clock by default.
More memory ranks also increases load on the memory controller which can results in lower achievable RAM clocks and higher RAM latencies. But in practice having 2 ranks per channel decreases RAM latencies under load due to parallelization of memory requests, and at least with Ryzen 4 ranks per channel typically decreases RAM latencies under load even further, resulting in higher CPU performance.
I'd go with two dual-rank 32GiB modules for a 64GiB system. I think all consumer 32GiB modules are dual rank anyway though. 4 dual-rank 16GiB modules might give better performance at the same clock speed but at least on Ryzen it probably won't clock as high, only 2667 MHz is guaranteed for such a configuration.

Thanks for details, very interesting.

I'm not sure what means "memory ranks" and "ranks per channel".
It has something to do with the number of DRAM chips mounted on a single memory module ?

Found this, but caused more confusion for me

RAM | Integral Memory

integralmemory.com

Also searched on specs page at manufacturer page.
Most don't have any mention about "ranks".
Example : SKU CMK64GX4M2C3200C16 , VENGEANCE® LPX 64GB (2 x 32GB) DDR4 DRAM 3200MHz C16 Memory Kit - Black

VENGEANCE® LPX 64GB (2 x 32GB) DDR4 DRAM 3200MHz C16 Memory Kit - Black

VENGEANCE LPX memory is designed for high-performance overclocking. The heatspreader is made of pure aluminum for faster heat dissipation, and the eight-layer PCB helps manage heat and provides superior overclocking headroom.

www.corsair.com

 

[cgeorgy]

@cgeorgy

You're looking to build what is basically a SFF PC and you want it low noise, but you're going to be potentially throwing some pretty heavy workloads at these proposed CPUs which may make them get a little toasty. The 10900k in particular could potentially eat up to 250W all by itself (albeit not for sustained operation; after 56s it's supposed to drop to 125W though some motherboards don't enforce that). The 5800X is going to top out at 142W sustained and the 5600X 88W. Of the three CPUs listed, the 5600X will easily be the easiest CPU to cool in your chosen enclosure. Are you sure 6c/12t will be enough for all your VMs? How many cores do you need per VM?

If you're going to step up to a 142W PPT CPU like a 5800x and if you want lots of cores for VMs, you really need to take the time to up your budget and get a 5900X or 5950X. The Intel chip will get you more cores than the 5600X and its availability is higher, but with its power-consumption you'll have to do a lot of work to either cool it or undervolt it.

Thanks for info.
I will go with Ryzen 7 5800x.
Now I have to decide what mobo.

Here is the continuation in Motherboards section :

Question - What mobo for Ryzen 7 5800x ?

Hello, I'm trying to decide about a mobo for Ryzen 7 5800x. I want reliability and stability more than high performance. I'm not a gamer, I just want to run a few VMs in parallel with Docker and various databases, do software development. No CPU intensive workloads running a lot of time...

forums.anandtech.com

 

[gdansk]

If you do go Ryzen, remember to enable SVM. Every AM4 motherboard I've had has it disabled by default

 

[blckgrffn]

I honestly feel that since you are trying to keep it quiet, just get a 5600x. You're paying a huge premium for cores that, to me, sound like they will be used infrequently.

In return you get a superbly balanced CPU w/regards to power usage at all times, including peak. This is, to me, like taking a ~60W bulb out of your mATX case.

Then you can use that extra $150 or so on a mediocre GPU. Once upon a time a 1650 or 5500xt would have been an easy recommendation, but now I fear you are looking at used GTX 1050's or new 1030's for Starcraft 2.

At least SC2 can run on darn near anything.

 

[DrMrLordX]

Thanks for info.
I will go with Ryzen 7 5800x.

Remember than the 5800X, 5900X, and 5950X are all 142W CPUs. The 5800X dissipates this heat from only one CCD, making it the hardest to cool during default operation of all the 5000-series CPUs.

 

[DrMrLordX]

Lol with the downvote. Sorry with the meta/fourth-wall here but when you get downvoted for posting something factually-correct, it makes you smile a little.

 

[VirtualLarry]

AM4 platform has support for SVM (Ryzen) as well as ECC (mobo-dependent, may need "Ryzen PRO" CPU, which is validated for ECC usage).

One thing that I feel that you should STRONGLY re-consider, is your choice of micro-ATX. If you are going to do hardware pass-through of GPUs to VMs, it really helps (so I heard from UnRAID forums, I haven't done it myself yet), to have MULTIPLE PCI-E x16 slots, as the pass-through is (I think?) slot-granular, for the most part. So if you want a GPU on the VM host, and then one for the VM, you're going to need two PCI-E x16 slots at a minimum, which means micro-ATX probably won't make the cut, or the GPUs will be horribly cramped and overheating.

 

[cgeorgy]

to have MULTIPLE PCI-E x16 slots, as the pass-through is (I think?) slot-granular, for the most part. So if you want a GPU on the VM host, and then one for the VM

So this means to have two video cards, right ?
This requires some sort of bridge between them such as SLI ?

 

[cgeorgy]

The 5800X dissipates this heat from only one CCD, making it the hardest to cool during default operation of all the 5000-series CPUs.

Not sure I understand your info.
Did some research in order to understand what is a CCD.
But how 5800X can dissipate more heat compared to a 5950X which is able to have more cores in full load at a given time ?

 

[VirtualLarry]

So this means to have two video cards, right ?
This requires some sort of bridge between them such as SLI ?

No need for a bridge, this isn't SLI. Both cards need a physical PCI-E x16 slot, although they can be wired as x1 or x4 electrically, mining still works for most algos with minimal PCI-E bus bandwidth.

 

[Martimus]

Thanks for details, very interesting.

I'm not sure what means "memory ranks" and "ranks per channel".
It has something to do with the number of DRAM chips mounted on a single memory module ?

Found this, but caused more confusion for me

RAM | Integral Memory

integralmemory.com

Also searched on specs page at manufacturer page.
Most don't have any mention about "ranks".
Example : SKU CMK64GX4M2C3200C16 , VENGEANCE® LPX 64GB (2 x 32GB) DDR4 DRAM 3200MHz C16 Memory Kit - Black

VENGEANCE® LPX 64GB (2 x 32GB) DDR4 DRAM 3200MHz C16 Memory Kit - Black

VENGEANCE LPX memory is designed for high-performance overclocking. The heatspreader is made of pure aluminum for faster heat dissipation, and the eight-layer PCB helps manage heat and provides superior overclocking headroom.

www.corsair.com

Ranks are sets of parallel memory chips that can either read or write during any particular clock. Having more than 1 rank per DIMM allows them to do both at the same time (one rank reading, while another rank writes) effectively increasing the bandwidth.

Here is a good video that explains it better than I can, by someone who understands it better than I do:

And here is one that is much better organized and easier to follow:

 

[DrMrLordX]

But how 5800X can dissipate more heat compared to a 5950X which is able to have more cores in full load at a given time ?

5800X and 5950X both top out at 142W stock power consumption. 5950X manages this by running lower voltages and lower all-core clockspeeds in intense MT workloads. When dissipating that much heat from one CCD (which means Core Complex Die), the 5800X achieves higher temperatures. The 5950X dissipates 142W from two dice.

 

[coercitiv]

But how 5800X can dissipate more heat compared to a 5950X which is able to have more cores in full load at a given time ?

The simplest explanation is 5950X dissipates the same amount of heat from twice the surface area. This means that heat dissipation per core is lower, and heat transfer towards the heatsink is higher than on 5800X.

 

[B-Riz]

When you get the 5800X, try your workloads with the CPU in ECO mode, is a BIOS setting that turns it into a 65 watt cpu, like running a 5700X; it will lower your MT performance a little bit, but not drastically.

From my testing, it drops (all cores loaded) MT clocks 2-300MHz depending, but temps were down to 60*C versus 80*C.

Single thread is not affected. So, if you just have VM's and containers open, but not sustained heavy CPU usage, I don't think ECO mode would negatively affect you.

 

[Mloot]

Another option might be to find a 3900x. I have seen a few for sale in the $350 territory, from folks that have upgraded to Ryzen 5xxx. If you need alot of cores but do not necessarily need the fastest cores, you can save a couple of hundred under what a similarly equipped 5900x would cost.

 

[B-Riz]

Another option might be to find a 3900x. I have seen a few for sale in the $350 territory, from folks that have upgraded to Ryzen 5xxx. If you need alot of cores but do not necessarily need the fastest cores, you can save a couple of hundred under what a similarly equipped 5900x would cost.

That was going to be my recommendation too, but, they play Starcraft 2, and Zen3 is just a beast at the game compared to Zen2.

https://www.reddit.com/r/Amd/comments/jp5qda

https://www.reddit.com/r/Amd/comments/jpie29

From 2nd link:

CPU/Mode	Min	Avg	Max
3600/4v4	3	43.385	188
5600X/4v4	37	199.843	674 (Really)

 

[Justinus]

5800X and 5950X both top out at 142W stock power consumption. 5950X manages this by running lower voltages and lower all-core clockspeeds in intense MT workloads. When dissipating that much heat from one CCD (which means Core Complex Die), the 5800X achieves higher temperatures. The 5950X dissipates 142W from two dice.

Lol I have no idea how you quoted this attributed to me but the forums alerted me as if it was actually my post you quoted...

 

[damian101]

Thanks for details, very interesting.

I'm not sure what means "memory ranks" and "ranks per channel".
It has something to do with the number of DRAM chips mounted on a single memory module ?

Found this, but caused more confusion for me

RAM | Integral Memory

integralmemory.com

Also searched on specs page at manufacturer page.
Most don't have any mention about "ranks".
Example : SKU CMK64GX4M2C3200C16 , VENGEANCE® LPX 64GB (2 x 32GB) DDR4 DRAM 3200MHz C16 Memory Kit - Black

VENGEANCE® LPX 64GB (2 x 32GB) DDR4 DRAM 3200MHz C16 Memory Kit - Black

VENGEANCE LPX memory is designed for high-performance overclocking. The heatspreader is made of pure aluminum for faster heat dissipation, and the eight-layer PCB helps manage heat and provides superior overclocking headroom.

www.corsair.com

I think in the consumer space there are pretty much only single-rank and dual-rank modules. For the memory controller handling one dual-rank module per channel is pretty much the same as handling two single-rank modules per channel. Unlike with multiple RAM channels having more ranks doesn't give you any raw bandwidth advantage, but latencies under load can be reduced as requests can happen in parallel.
Manufacturers very often don't list the number of ranks per module in the specs. With bigger capacity you can expect more ranks, I think for DDR4 32 GiB is pretty much always dual-rank, 16GiB mostly dual-rank, 8GiB mostly single rank and 4GiB pretty much always single-rank.
Sometimes manufacturers will list RAM as single-side or dual-side, which doesn't necessarily coincide with single-rank and dual-rank, at least not when using the terms correctly.

 

[Ratman6161]

I've been running rigs like this for years and will point out something others have not mentioned. My experience is that with desktop hardware the biggest bottleneck is storage. In the earlier days when I was still using spinning disks I found that more than one vm running on the same disk had everything slowing to a crawl. I resorted to having multiple disks and spreading the VMs out on multiple disks. Now, even with SSDs I've got multiple SSDs and try not to put two VMs that routinely run at the same time on the same drives.

I'd also say that having lots of RAM is also more important than the CPU. IMO this is a case where quantity counts more than performance. VMWARE will let me over provision RAM but if the VMs demand more than the system has, once again everything will crawl.

Depending on just what the various VMs are doing, the CPU may matter a lot less than you think. A lot of the time most 9f my VMs are sitting there at idle waiting for input. I'd vote for more slower cores over a smaller number of faster cores.

You could also consider a separate system to host VMs.